Build for the Models Six Months Out
Ship for the AI capabilities that will land in ~6 months, not the ones available today
- Difficulty
- Advanced
- Time to result
- ~months to results
- Steps
- 4
- Confidence
- 88%
AI capability is improving on a roughly six-month cadence. Optimizing hard for what today's models can do is a trap — the modality you perfect (e.g. autocomplete) can be obsoleted by the next model generation. Instead, form a conviction about the trajectory, build the product the next generation of models will unlock, and ship it as those models land. Replit began building its agent when models literally could not yet do it, betting reasoning would arrive in six months — and it did.
Origin
Amjad Masad, describing how Replit timed its shift from autocomplete to agentic coding; he invokes Ray Kurzweil's point that humans systematically underestimate exponentials.
Core principles
- 01AI capability improves on a roughly six-month cadence (Sonnet V1 → V2 as the reference points)
- 02Exponentials are hard for humans to grasp — you must consciously overcorrect toward the future
- 03Optimizing for today's model locks you into a modality that the next model will make obsolete
- 04The winning modality can change entirely (autocomplete → in-environment chat/agent)
- 05You must build the product before the model that runs it exists
How to run it
- 1
Convince yourself the trajectory is real
Explicitly decide whether AI is on a sustained exponential. This conviction is the hard part and it feels absurd to an engineering mind, but every downstream bet depends on it.
Pro tip Use Kurzweil's framing — humans instinctively underestimate exponential curves — as a deliberate counterweight to your intuition.
Watch out If you can't honestly convince yourself the curve holds, don't make this bet — it only pays off if the trajectory is real.
- 2
Identify the modality the next models unlock
Ask what product form becomes possible once models are ~6 months better, not what's optimal today. For Replit the answer moved from autocomplete to chatting inside the programming environment with an agent that builds for you.
Watch out Don't over-invest in perfecting the current modality — Replit trained and heavily optimized its own autocomplete models in 2022, then had to abandon that modality.
- 3
Build ahead of the capability
Start building for models landing in six months even while today's models can't do it, so the product is ready the moment the capability arrives.
- 4
Ship as the models land and re-cadence
Release when the capable model arrives, then repeat on the next six-month cycle as reasoning and capability compound.
Pro tip Keep the product architecture ready to swap in a stronger model — Replit saw reasoning 'improve so much' just by switching model versions.
In the wild
In 2022 Replit built and optimized its own autocomplete/copilot-style models. Masad recognised that modality was becoming the wrong one and, a year before the capability existed, directed the team to build the agent for models that would land in six months. Six months later the models capable of the needed reasoning arrived.
→ Replit shipped an agent that builds and maintains full-stack apps, riding the Sonnet V1→V2 improvement curve rather than being trapped in the obsoleted autocomplete modality.
Common mistakes
Optimizing for today's model
Perfecting a product around current capabilities (like Replit's heavily-tuned 2022 autocomplete) wastes effort on a modality the next model generation renders obsolete.
Trusting your linear intuition about progress
Humans underestimate exponentials, so an un-corrected gut feel will consistently tell you the future is further away than it is, causing you to build too conservatively.
Is it for you?
Best for
Founders and product leaders building in fast-moving AI categories who must decide what to build before the enabling technology exists
Not ideal for
Stable, slow-moving domains where capability isn't compounding and building ahead just ships something nobody can use yet
From the transcript
“we were like okay we're going to build for the models that are landing in six months and it truly like six months later the…”
“it's easy to build and optimize for today you know in in 22 we built like um you know coped like thing and autocomplete we…”
“Rikers well this like you know futurist um you know talks about how exponentials are really hard for humans to to grasp”
From the episode
Behind the product: Replit
Amjad Masad (co-founder and CEO)